inference-engine/thirdparty/clDNN/src/gpu/max_unpooling_gpu.cpp

   1 /*
   2 // Copyright (c) 2018 Intel Corporation
   3 //
   4 // Licensed under the Apache License, Version 2.0 (the "License");
   5 // you may not use this file except in compliance with the License.
   6 // You may obtain a copy of the License at
   7 //
   8 //      http://www.apache.org/licenses/LICENSE-2.0
   9 //
  10 // Unless required by applicable law or agreed to in writing, software
  11 // distributed under the License is distributed on an "AS IS" BASIS,
  12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13 // See the License for the specific language governing permissions and
  14 // limitations under the License.
  15 */
  16
  17 #include "max_unpooling_inst.h"
  18 #include "primitive_gpu_base.h"
  19 #include "implementation_map.h"
  20 #include "error_handler.h"
  21 #include "network_impl.h"
  22 #include "kernel_selector_helper.h"
  23 #include "max_unpooling/max_unpooling_kernel_selector.h"
  24 #include "max_unpooling/max_unpooling_kernel_base.h"
  25 #include <vector>
  26
  27 namespace cldnn {
  28 namespace gpu {
  29
  30 struct max_unpooling_gpu : typed_primitive_gpu_impl<max_unpooling> {
  31     using parent = typed_primitive_gpu_impl<max_unpooling>;
  32     using parent::parent;
  33
  34 protected:
  35     kernel::kernel_arguments_data get_arguments(typed_primitive_inst<max_unpooling>& instance,
  36                                                         int32_t split) const override {
  37         kernel::kernel_arguments_data args = parent::get_arguments(instance, split);
  38         args.inputs.push_back((memory_impl::cptr) &instance.dep_memory(1));
  39         return args;
  40     }
  41
  42 public:
  43     event_impl::ptr execute_impl(const std::vector<event_impl::ptr>& events, max_unpooling_inst& instance) override {
  44         // clear output buffer
  45         std::vector<event_impl::ptr> tmp_events(events);
  46         auto ev = instance.get_network().get_engine().create_user_event(instance.get_network().get_stream_id(), false);
  47         instance.output_memory().fill(0, ev);
  48         tmp_events.push_back(ev);
  49         return parent::execute_impl(tmp_events, instance);
  50     }
  51
  52     static primitive_impl* create(const max_unpooling_node& arg) {
  53         auto max_unpooling_params = get_default_params<kernel_selector::max_unpooling_params>(arg);
  54         auto max_unpooling_optional_params =
  55             get_default_optional_params<kernel_selector::max_unpooling_optional_params>(arg.get_program());
  56
  57         max_unpooling_params.inputs.push_back(convert_data_tensor(arg.argmax().get_output_layout()));
  58
  59         auto& kernel_selector = kernel_selector::max_unpooling_kernel_selector::Instance();
  60         auto best_kernels = kernel_selector.GetBestKernels(max_unpooling_params, max_unpooling_optional_params);
  61
  62         CLDNN_ERROR_BOOL(arg.id(),
  63                          "Best_kernel.empty()",
  64                          best_kernels.empty(),
  65                          "Cannot find a proper kernel with this arguments");
  66
  67         auto max_unpool = new max_unpooling_gpu(arg, best_kernels[0]);
  68
  69         return max_unpool;
  70     }
  71 };
  72
  73 namespace {
  74 struct attach {
  75     attach() {
  76         implementation_map<max_unpooling>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::yxfb),
  77                                                max_unpooling_gpu::create);
  78         implementation_map<max_unpooling>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::yxfb),
  79                                                max_unpooling_gpu::create);
  80         implementation_map<max_unpooling>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::bfyx),
  81                                                max_unpooling_gpu::create);
  82         implementation_map<max_unpooling>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::bfyx),
  83                                                max_unpooling_gpu::create);
  84         implementation_map<max_unpooling>::add(std::make_tuple(engine_types::ocl, data_types::i8, format::bfyx),
  85                                                max_unpooling_gpu::create);
  86         implementation_map<max_unpooling>::add(std::make_tuple(engine_types::ocl, data_types::i8, format::yxfb),
  87                                                max_unpooling_gpu::create);
  88         implementation_map<max_unpooling>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::byxf),
  89                                                max_unpooling_gpu::create);
  90         implementation_map<max_unpooling>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::byxf),
  91                                                max_unpooling_gpu::create);
  92         implementation_map<max_unpooling>::add(std::make_tuple(engine_types::ocl, data_types::i8, format::byxf),
  93                                                max_unpooling_gpu::create);
  94     }
  95     ~attach() {}
  96 };
  97 attach attach_impl;
  98 }  // namespace
  99 }  // namespace gpu
 100 }  // namespace cldnn